/ Anima_nomade, stock.adobe.com
LONDON – Machine learning-based software has detected Alzheimer’s disease with nearly 100% certainty using conventional 1.5 Tesla MRI. The results were inside Communications medicine (2022; DOI: 10.1038 / s43856-022-00133-4).
If dementia is caused by deposits of beta amyloids and tau fibrils, the histological characteristics of Alzheimer’s disease can only be determined after death for a long time. It is now possible to make the deposits visible using positron emission tomography (PET) or to detect their degradation products in the CSF.
However, a PET scan is expensive and only available in a few centers; lumbar puncture is an invasive and painful examination. An alternative would be magnetic resonance imaging (MRI). However, the changes there, such as atrophy in the two hippocampi, are subtle. MRI therefore has not so far been suitable for diagnosis.
This may change with new software being presented by a team led by Eric Aboagye of Imperial College London. The software, which is based on machine learning methodology, analyzes no less than 29,520 different morphological-functional characteristics in 115 brain regions.
It was trained and validated on participants in the Alzheimer’s Disease Neuroimaging Initiative, comparing uptake with healthy controls and patients with other neurological disorders including frontotemporal dementia and Parkinson’s disease. Finally, 83 patients from ongoing operations at Imperial College London were also examined.
According to Aboagye, the “Alzheimer’s Predictive Vector” can reliably distinguish between people with and without Alzheimer’s related diseases: accuracy was 98% in the “Alzheimer’s Disease Neuroimaging Initiative” and 81% in an external validation.
It far surpassed standard MRI of hippocampal atrophy (26% accuracy) and also eclipsed CSF testing for beta-amyloid (62% accuracy).
A disadvantage of the exam is the high expenditure of time. They are 10 to 12 hours for individual patients. The software is also trained on MRI machines with a strength of 1.5 Tesla. Paradoxically, with more powerful devices, it was more error prone. © rme / aerzteblatt.de